Announcing the ICLR 2026 Outstanding Papers

Guest Post by the ICLR 2026 Outstanding Paper Committee (Gautam Kamath (chair), Yoav Artzi, Emma Brunskill, Murat Erdogdu, Pang Wei Koh, Yong Jae Lee, Subhransu Maji, Doina Precup, Ohad Shamir, Guy Van den Broeck, Kilian Weinberger, Luke Zettlemoyer)

We are pleased to recognize the following two Outstanding Papers and one Honorable Mention for ICLR 2026.

Outstanding Papers

Transformers are Inherently Succinct, by Pascal Bergsträßer, Ryan Cotterell, Anthony Widjaja Lin

This theoretical work proposes a new perspective to explain the power of the Transformer architecture: namely, how succinctly it can encode some concepts compared to alternative models such as RNNs. Notwithstanding critiques, the paper’s strong conceptual message was intriguing to the committee and other experts. This work may stimulate additional theoretical and empirical investigation into succinctness of concept representation by transformers and other architectures.

LLMs Get Lost In Multi-Turn Conversation, by Philippe Laban, Hiroaki Hayashi, Yingbo Zhou, Jennifer Neville

There is a dissonant gap between most of the data used for LLM training and how they are deployed: the data is largely text completion or single turn, but the deployment is inherently in multi-turn settings. Robustly evaluating and training multi-turn capabilities remains a complex challenge. This paper designs a scalable method to evaluate multi-turn capabilities, and measures a marked decrease in LLM aptitude and reliability in the all too common case where interactions are multi-turn and involve underspecified instructions. The work demonstrates exceptional experimental design and methodology, and the findings are fresh and interesting, particularly for an important setting that more closely reflects real-world usage. Although concerns were discussed around the use of dated models, the committee considers the conclusions and method to remain relevant to state-of-the-art models. This work serves to highlight a problem which state-of-the-art models are optimized to solve, and provides a viable and scalable diagnosis.

Honorable Mention

The Polar Express: Optimal Matrix Sign Methods and their Application to the Muon Algorithm, by Noah Amsel, David Persson, Christopher Musco, Robert M. Gower

In this work, the authors use approximation theory to design optimal polynomial approximations for the polar decomposition, employed in the popular Muon optimizer. A particular focus is placed on approximation for modern deep learning settings, where computation is performed on GPUs and potentially with low-precision arithmetic. While empirical improvements were at times modest, the principled approach to (and general methodology for) improving one of the most popular optimizers was appreciated. 

Selection Process

Given the significant value associated with awards in our community, the committee followed a thorough and rigorous selection process, inspired by that of TMLR. This process is centered around the principle that, while the committee has significant and broad expertise, the individuals best suited to evaluate the contributions of a paper are experts in that specific area. In contrast to TMLR’s process (which is not time-sensitive), we necessarily had to work on a shorter timeframe, and the selection was conducted over five weeks in March and April. Details of the selection process follow. 

Selection was performed in three phases. In the first phase, a “longlist” of award candidates was identified by the Program Chairs. This includes 20 papers which were flagged as potentially award-worthy by the paper’s Area Chair, and 17 papers which had very high scores from the review process. This produced a long-list of 36 papers for consideration by the committee (one paper was flagged by both mechanisms). 

In the second phase, these 36 papers were divided amongst the 12 committee members (3 papers per committee member). The job of each committee member was to advance 0 to 1 papers from their batch (exceptionally 2). Committee members were directed to form their opinion on each paper using a variety of signals: their own reading of the paper, reading the reviews, and, crucially, soliciting the opinion of multiple experts in the paper’s area, including the Area Chair, Senior Area Chair, and well as 1-3 other researchers who work as closely to the paper’s results as possible. Conflicts of Interest were respected during this stage: no committee member was assigned to a paper that they have a strong conflict with the authors, and feedback from external experts during this phase was only visible to the committee member and the chair (who was not able to see comments related to papers that he had a conflict with). Overall, during this phase, 5 committee members chose to advance 1 paper, and 7 committee members chose to advance 0 papers, leading to a shortlist of 5 papers. 

In the third and final phase, all committee members were to evaluate the 5 papers in the shortlist, again, using the same signals as before (their own reading, the reviews, as well as external expert evaluations). They were then instructed to vote for the type of recognition each paper should receive (Outstanding Paper, Honorable Mention, or no recognition). The program chairs looked at the voting results, and chose to recognize two papers as Outstanding Papers, one paper as Honorable Mention, and the two remaining papers with no additional recognition. Conflicts were again respected during this phase: one committee member was excused from deliberation and voting, and the committee chair (who has a conflict with one recognized paper) exclusively shepherded this phase by collecting votes of committee members: he did not vote himself, and the final decision of what accolades should be assigned was performed by the Program Chairs. 

We acknowledge that, like all award selection procedures, this process is inherently subjective and susceptible to overlooking excellent work. Indeed, some of the best research can be rather controversial to experts in the area. We look forward to 2036, when the community will revisit the papers at ICLR 2026 to see which have stood the test of time. 

Some miscellanea:

• Towards the start of the process, there was discussion amongst the committee members regarding how many papers ought to be recognized. There was generally an appetite for recognizing more than one or two papers, given the significant number of submissions. Nonetheless, over the course of the process, the committee converged to recognition of a relatively small number of papers (three). 
• In the middle of the committee’s deliberation, one committee member suggested that the selection process ought to be blind to the identity of the paper’s authors. While we generally agree that this should not play a role in award selection, logistically, it appears difficult to enforce for a number of reasons (e.g., award selection is performed after accepted papers are announced, conflicts of interest may be difficult to respect, etc.). Future award committees may consider whether it is feasible to mask the identity of the authors.